Global Forecasting of Tropical Cyclone Intensity Using Neural Weather Models

TL;DR

This work addresses the challenge of predicting tropical cyclone intensity with globally available neural weather models that operate at coarse resolutions. By post-processing NeWM outputs from PanguWeather and FourCastNet v2 with a tracking-independent framework, the authors obtain accurate probabilistic forecasts of $\Delta V_{max}$ and $\Delta P_{min}$ up to $\tau=168$ hours using ground-truth data from IBTrACS and ERA5 as input. The study compares linear models, ANNs, CNNs, and UNets, finding that simple linear post-processing can achieve competitive probabilistic skill, while CNNs excel at short leads but struggle later; masking helps, and the overall approach democratizes global TC-intensity forecasting by reducing computational barriers. The results underscore the value of integrating global NeWMs with lightweight post-processing to deliver accessible, end-to-end TC intensity forecasts, with clear avenues for incorporating oceanic data and broader datasets in future work.

Abstract

Numerical Weather Prediction (NWP) models that integrate coupled physical equations forward in time are the traditional tools for simulating atmospheric processes and forecasting weather. With recent advancements in deep learning, AI-based Weather Prediction models that rely on neural network architectures$\unicode{x2013}$Neural Weather Models (NeWMs)$\unicode{x2013}$have emerged as competent medium-range NWP emulators, with performances that compare favorably to state-of-the-art NWP models. However, they are commonly trained on reanalyses with limited spatial resolution (e.g., 0.25° horizontal grid spacing), which smooths out key features of weather systems. For example, tropical cyclones (TCs)$\unicode{x2013}$among the most impactful weather events due to their devastating effects on human activities$\unicode{x2013}$are challenging to forecast, as extrema are smoothed in deterministic forecasts at 0.25° resolution. To address this, we use our best observational estimates of wind gusts and minimum sea level pressure to train a hierarchy of post-processing models on NeWM outputs. Applied to Pangu-Weather and FourCastNet v2, the post-processing models produce accurate and reliable forecasts of TC intensity up to five days ahead. Our post-processing algorithm is tracking-independent, preventing full misses, and we demonstrate that even linear models extract predictive information from NeWM outputs beyond what is encoded in their initial conditions. While spatial masking improves probabilistic forecast consistency, we do not find clear advantages of convolutional architectures over simple multilayer perceptrons for our NeWM post-processing purposes. Overall, by combining the efficiency of NeWMs with a lightweight, tracking-independent postprocessing framework, our approach improves the accessibility of global TC intensity forecasts, marking a step toward their democratization.

Figures (8)

• Figure 1: We propose post-processing NeWM forecasts to improve the prediction of TC intensity. Our postprocessing pipeline includes three steps: first, we generate forecast fields using neural weather models. Second, we clip the global fields produced by the models to a bounded region, using the reported location from IBTrACS at the initial time of forecast. Finally, we try to match observation targets using a deterministic or a probabilistic post-processing model.
• Figure 2: Training Set Quantiles for (a) Zonal Displacement and (b) Meridional Displacement
• Figure 3: Masked wind magnitude field for the PanguWeather forecast at initial time 2020-01-11 00h00 (associated with TC Claudia) for different lead times. The black plus ($+$) symbols indicate the position of TC Claudia at the given lead time according to IBTrACS.
• Figure 4: Wind Magnitude feature pixel-wide distribution on the training and validation set (unmasked inputs). For all feature distributions, refer to Figure S-2.
• Figure 5: Training curves for (a) comparing across algorithms trained on PanguWeather outputs and (b) comparing algorithm performance when trained on PanguWeather (P) and FourCastNet v2 (F) outputs. All inputs were masked, T denotes Training, V denotes Validation.